PUEB-DimASR at SemEval-2026 Task 3: Escaping the Mean Regression Trap with Graph-Enhanced Transformers for Dimensional Aspect-Based Sentiment Regression
Year: 2026
Type: conference-paper
Abstract: The DimABSA shared task aims to combine dimensional analysis with aspect-based sentiment analysis (ABSA).It addresses the lack of continuous sentiment representation, as opposed to categorical labels (e.g., positive, negative, or neutral), and enriches it with an assessment of arousal.Our team's PUEB-DimASR investigates the "mean-regression trap" -the tendency of the MSE loss in high-dimensional sentiment tasks to over-predict values closer to the global mean.Within the paper we propose a two-step architecture.First, we enhance baseline transformers with graph convolutional networks (GCN) to capture syntactic aspectsentiment dependencies.Second, we evaluate and recommend a hybrid loss function that combines mean squared error (MSE) and concordance correlation coefficient (CCC).Our proposed GCN-deBERTa model consistently outperforms the baseline across six target languages.While MSE loss yields the best RMSE scores for English (0.876) and Chinese (0.546), it introduces significant variance collapse, which we successfully mitigated using the hybrid loss, achieving near-perfect distributional alignment (99.6%).Additionally, our model trained with the hybrid loss achieved the best RMSE scores for Russian (1.136), Tatar (1.207), and Ukrainian (1.178).
Authors: Oskar Riewe-Perła, Agata Filipowska
Institution: Poznań University of Economics and Business
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166872293